Same Prompt, Different Test: Using AI Agent Workspaces to Make Testing Repeatable

The Slack number is the part that frames the whole problem well, only about 20 percent of runs following the same sequence. Without that, repeatability sounds like a nice extra instead of the core requirement for regression tests and audits. The three approval gates are the detail I would keep too, especially the one before the write to qTest, since a bad decision saved into a workflow repeats on every run. The point about humans deciding whether a test checks the right thing is the honest limit of the whole approach, and it is good to see it stated plainly.